Careful with Causal Inference

Richard Emsley
Professor of Medical Statistics and Trials Methodology
Department of Biostatistics & Health Informatics
Institute of Psychiatry, Psychology & Neuroscience
 Correlation and causation

“Some scientists are reluctant to speak so

blatantly about cause and effect, but in
statements of hypothesis and in describing
study objectives such boldness serves to
keep the real goal firmly in focus and is
therefore highly preferable to insipid
statements about ‘association’ instead of Messerli F (2012), NEJM, 367:16
‘causation’

Rothman (1986), Modern Epidemiology

2
 Overview
1. What is causal inference?

2. How does causal inference help?

➢ Better causal questions
➢ Better confounding control

3. How can we be more careful with causal inference?

➢ Or should we be more explicit?

3
If there is a significant correlation between two
variables R and Y, then either:
1. R causes Y R Y

2. Y causes R R Y

R Y
3. R and Y share a common cause X

R Y
4. R and Y are conditioned on a
common descendent Z
Z
4
 The general principle of causal inference
• Statistical models can only tell us about association between
two variables (say R and Y)

• The aim of causal inference is to infer whether this association

can be given a causal interpretation (e.g. R causes Y) by:
– defining the causal estimands
– being explicit about the assumptions being made
– thinking about other possible explanations for observed effects,
especially confounding.

• There are now many, many methods purporting to give

causally valid solutions to this problem; this session only gives
an overview of some of these

5
 A brief history of causal inference (1)

• Neyman (1923) and Fisher

(1925) discussed the
potential yield to be gained
from agricultural plots under
different experimental
exposures.

• First introduction of the Ronald Fisher Jerzy Neyman

(1890-1962) (1894-1981)
concept of random
allocation as an
experimental design.

6
 A brief history of causal inference (2)

• This was formalised statistically for both

randomised and non-randomised studies
many years later.
– Potential outcomes
– Rubin Causal Model (Holland 1986)

• Rubin DB (1974). Estimating causal effects

of treatments in randomized and
nonrandomized studies. Journal of
Educational Psychology 66(5), 688-701.
• Rosenbaum PR and Rubin DB (1983). The
central role of the propensity score in Don Rubin
observational studies for causal effects.
Biometrika 70(1), pp41-55.

7
 A brief history of causal inference (3)

• Extended the potential outcomes

framework to longitudinal setting
(repeated measures).

• This required a new methodology

for estimating parameters using
semi-parametric theory: the “G-
family”
Jamie Robins

• Uses terminology ‘counterfactuals’

rather than potential outcomes.

8
 A brief history of causal inference (4)

• Developed a theory of causal and

counterfactual inference based on
graphical models and probabilistic
reasoning.

• Derived a new method

for determining relations Judea Pearl
between variables, known
as ‘do-calculus’.

• Explores the link between

counterfactuals and
non-parametric structural
equation models.

9
 A brief history of causal inference (5)
• There is a group who argue
against using the counterfactuals
or potential outcomes framework.

• Dawid and colleagues propose

for methods for causal inference
without counterfactuals, L-R: Carlo
mainly using decision theory, Berzuini, Phil
graphical models and Dawid, Vanessa
Didelez
stochastic modelling.

10
 Is the terminology important?

“Personally I see the different formalisms as different

‘languages’. The French language may be best for
making love whereas the Italian may be suitable for
singing, but both are indeed possible…”

Lauritzen: Scandinavian Journal of Statistics 2004 Vol. 31 p189

11
Fundamental concept of causal inference

Receive treatment Receive control

Measure outcome Measure outcome

Comparison of outcomes gives an
individual treatment effect 12
Fundamental concept of causal inference

Receive treatment Receive control

Measure outcome Measure outcome

Comparison of these outcomes will not give an individual
treatment effect 13
Fundamental concept of causal inference

Receive treatment Receive control

Measure outcome Measure outcome

Comparison of average outcomes defines the

average treatment effect
Fundamental concept of causal inference

Receive treatment Receive control

Measure outcome Measure outcome

Comparison of average outcomes estimates the

average treatment effect
 Directed Acyclic Graphs

• A causal effect of one variable on another (R: treatment, Y:

outcome) is shown as:

R Y

• If Y is continuous variable then we could estimate the effect using

linear regression:
y𝑖 = 𝛽0 + 𝛽1 r𝑖 + 𝜀𝑖 , 𝜀𝑖 ~𝑁(0, 𝜎 2 )

• A DAG must not be cyclic, i.e. starting at treatment we should be

able to get back to it

R Y
 Types of variable in DAGs

• Confounding (X)
• Mediation (M)
• Partitions effects into direct and indirect effects
• Colliders (Z) – biases association between their parents
X

R M Y

Z
 Overview
1. What is causal inference?

2. How does causal inference help?

➢ Better causal questions
➢ Better confounding control

3. How can we be more careful with causal inference?

➢ Or should we be more explicit?

18
 When can ATE be estimated?
• We wish to evaluate the effects of receiving a treatment compared
to a suitably defined control condition
• Thus we want to use a sample of subjects from a relevant target
population to compare outcomes between a treated group and a
control group
• When can we do this without running into problems?
– Randomised controlled trial (RCT): Participants are randomised to two
arms (experimental treatment and control)
• YES - provided participants adhere to their allocated treatment.
– Observational study: Compares subjects receiving the experimental
treatment with subjects under the control condition
• SOMETIMES - only if variables that drive treatment group selection
have been measured and accounted for appropriately

19
 What are we estimating in trials?
• Interested in various measures of effect
– Effectiveness - the benefit of a treatment policy
– Efficacy - the benefit of actually receiving treatment

• ITT measures effectiveness as implemented in a given trial

• What is the effectiveness of offering the intervention?

• It tells us whether randomising the treatment works

– On average, not for an individual patient!
– Regardless of whether you receive the treatment or not!

20
 Target mechanisms

Mechanism

Intervention Symptoms

• Target intermediate variables:

– Some treatments target a particular intermediate variable in order to bring
about change in a clinical outcome.
• Motivational interviewing → substance use → symptoms
• Cognitive behaviour therapy → thinking → symptoms
• Beta blockers → blood pressure → stroke risk
• Sleep intervention → sleep → cognition

• An explanatory analysis of a trial would seek to establish that this is indeed the
case using mediation analysis; i.e. assess the mediated path.
21
Mediation analysis and causal inference…

“Mediation analysis is a form of causal

analysis…all too often persons conducting
mediational analysis either do not realize that
they are conducting causal analyses or they
fail to justify the assumptions that they have
made in their casual model.”
David Kenny (2008), Reflections on Mediation, Organizational
Research Methods.

22
 The basic underlying problem:
estimating valid causal effects
U – the unmeasured confounders
error
U

Mediator

Random error
Outcomes
allocation

Covariates

23
 Example: simple mediation analysis

e1

0.25 X

M 0.25 0.25
0.25

0.25
R Y e2

Total effect = direct effect + indirect effect

= 0.25 + 0.25*0.25
= 0.25 + 0.0625
24
 Example: regression approach stage 1

. regress y r

Source SS df MS Number of obs = 2,000

F(1, 1998) = 289.30
Model 49.0074641 1 49.0074641 Prob > F = 0.0000
Residual 338.463931 1,998 .169401367 R-squared = 0.1265
Adj R-squared = 0.1260
Total 387.471395 1,999 .193832614 Root MSE = .41158

y Coef. Std. Err. t P>|t| [95% Conf. Interval]

r .3130773 .0184068 17.01 0.000 .2769787 .3491759

_cons .3082655 .012983 23.74 0.000 .2828039 .3337272

25
 Example: regression approach stage 2

. regress m r

Source SS df MS Number of obs = 2,000

F(1, 1998) = 260.01
Model 31.8440108 1 31.8440108 Prob > F = 0.0000
Residual 244.698233 1,998 .122471588 R-squared = 0.1152
Adj R-squared = 0.1147
Total 276.542244 1,999 .138340292 Root MSE = .34996

m Coef. Std. Err. t P>|t| [95% Conf. Interval]

r .252368 .0156509 16.12 0.000 .2216743 .2830617

_cons .2383794 .0110391 21.59 0.000 .21673 .2600288

26
 Example: regression approach stage 3

. regress y m r

Source SS df MS Number of obs = 2,000

F(2, 1997) = 1002.47
Model 194.120376 2 97.0601879 Prob > F = 0.0000
Residual 193.351019 1,997 .096820741 R-squared = 0.5010
Adj R-squared = 0.5005
Total 387.471395 1,999 .193832614 Root MSE = .31116

y Coef. Std. Err. t P>|t| [95% Conf. Interval]

m .7700831 .0198916 38.71 0.000 .7310728 .8090935

r .1187329 .0147935 8.03 0.000 .0897207 .1477452
_cons .1246936 .0109006 11.44 0.000 .1033158 .1460713

27
 Example: regression approach stage 3

. regress y m r x

Source SS df MS Number of obs = 2,000

F(3, 1996) = 1359.50
Model 260.153288 3 86.7177628 Prob > F = 0.0000
Residual 127.318106 1,996 .063786626 R-squared = 0.6714
Adj R-squared = 0.6709
Total 387.471395 1,999 .193832614 Root MSE = .25256

y Coef. Std. Err. t P>|t| [95% Conf. Interval]

m .241628 .0230313 10.49 0.000 .1964601 .2867958

r .2540715 .0127229 19.97 0.000 .2291199 .279023
x .2605752 .0080987 32.17 0.000 .2446924 .2764581
_cons .2510649 .0096803 25.94 0.000 .2320804 .2700495

28
 Example: with measurement error

. regress y m_star r x

Source SS df MS Number of obs = 2,000

F(3, 1996) = 1277.52
Model 254.780967 3 84.9269891 Prob > F = 0.0000
Residual 132.690428 1,996 .06647817 R-squared = 0.6575
Adj R-squared = 0.6570
Total 387.471395 1,999 .193832614 Root MSE = .25783

y Coef. Std. Err. t P>|t| [95% Conf. Interval]

m_star .052463 .0105353 4.98 0.000 .0318016 .0731244

r .3014506 .0118715 25.39 0.000 .2781688 .3247324
x .3077627 .0063906 48.16 0.000 .2952298 .3202957
_cons .2973219 .0084511 35.18 0.000 .280748 .3138957

29
 “Nuisance” mediators
• Variables measured post-randomisation that we may wish to
rule out having a mediated effect - essentially we want to
estimate the residual direct effects and find a small indirect
effect

• Use of concomitant medication or interventions in treatment

as usual

Antidepressant
Use
Cognitive therapy +
antidepressant Depression
Vs. Score
Antidepressant only
30
 Example: HIV and condom use
• Diaphragm and lubricant gel for prevention of HIV acquisition
in southern African women: a randomised controlled trial

• Padian et al. (2007) Lancet 370(9583):251-61.

Diaphragm + gel +
condoms
HIV infection
Vs
Condoms alone

31
Example: ITT effect on HIV incidence

• Overall HIV incidence was 4.0%

per 100 woman-years: 4.1% in
the intervention group (n=2472)
and 3.9% in the control group
(n=2476), corresponding to a
relative hazard of 1.05 (95% CI
0.84-1.32, intention-to-treat
analysis)
• Padian et al (2007)

32
 Example: ITT effect on condom use

• The proportion of women using condoms was significantly lower in the

intervention than in the control group
(54% vs 85% of visits, p<0.0001).

• Padian et al (2007)

33
 Example: mediation question
• “Shelton and Stein also ask for a disentanglement of the
separate effects of condom and diaphragm use on incidence
of HIV infection in the trial, part of which necessarily involves
an estimation of the independent effect of condom use. We
agree that such analyses are important additions to basic
intention-to-treat results, and in fact, we prespecified
appropriate methods to address these issues in our analytical
plan and included results in the submitted paper. Regrettably,
we were directed not to include these findings by both a
referee and an editor of the original article.”

• Jewell et al (2007), The Lancet, 370(9602):1823-1824

34
 Example: mediation hypothesis
• A case for controlled direct effects…

• What is the direct effect of randomisation to diaphragm use

on HIV infection if everyone in the population used
condoms?

• What is the direct effect of randomisation to diaphragm use

on HIV infection if no-one in the population used condoms?
Condom use

Diaphragm + gel + condoms

Vs HIV infection
Condoms alone
35
 Example: mediation analysis
• The estimated relative risk of HIV infection for assignment to
the intervention versus control group, had all participants
been constrained to always use condoms, was 0.96 (95% CI
0.59–1.45).

• By contrast, the estimated analogous relative risk of HIV

infection had all participants never used condoms was 0.59
(0.26–4.56).

• Jewell et al (2007), The Lancet, 370(9602):1823-1824

36
 Assumptions for identification
• Controlled direct effects require:
– A1: no unmeasured R-Y confounding (U1);
– A2: no unmeasured M-Y confounding (U2).

M U2

R Y
X
U1
37
 M-bias graph

G B
X

D Y

38
 M-bias graph
• What are all the paths from D to Y?
G B
X
• What is the consequence of adjusting
for X in estimating the effect of D on Y? D Y

• What variables do we need to adjust for to estimate the effect

of D on Y?

39
 A realistically complex DAG
• To estimate the effect of obesity on PE (Pearce and Lawlor, IJE, 2017)

• Unblocked paths:
• Obesity – Smoking – SEP - Age at pregnancy - PE
• Obesity – SEP – Age at pregnancy - PE
• Obesity – Smoking – Age at pregnancy - PE
• Obesity – Smoking – PE

• Adjusting for age at pregnancy and Smoking is sufficient

 Confounding
• Confounding is a causal concept G B
– Confounding of which effect? X
– Cannot causally interpret the
parameters of confounders
D Y
• Association does not require any confounding adjustment
– simply compute it from the data

• Prognostic models do not require confounding adjustment

– Include X and B

• Propensity score models should not include predictors of

exposure (variable G)
41
 Non-compliance in trials

B
X

R D 𝛽
Y

• R = randomisation
• D = treatment received
• Y = outcome Effect of R on Y
𝛽=
• X = confounder Effect of R on D
• B = prognostic variable

42
 Mendelian randomisation

B
X

G D 𝛽
Y

• G = gene
• D = treatment received
• Y = outcome Effect of G on Y
𝛽=
• X = confounder Effect of G on D
• B = prognostic variable

43
 Overview
1. What is causal inference?

2. How does causal inference help?

➢ Better causal questions
➢ Better confounding control

3. How can we be more careful with causal inference?

➢ Or should we be more explicit?

44
 Should we be careful with the C-Word?
• Miguel Hernán (2018), American Journal of Public Health

45
 Observational studies and causality
• In observational studies, we want to know about variables
that can be modified or manipulated

• Define the causal effect in the population as the causal effect

that would have been observed in a hypothetical trial

• Your observational effect estimate may be seriously

confounded…but we know that

• The goal is causal…the analysis is associational

• This is true of randomised trials as well!

46
 When can we use the words “causal effect”?

• Title

• Introduction

• Methods

• Results

• Discussion 47
 Hayes and Rockwood (2017)
“There are some hardliners who say that to claim the existence of
cause-effect relationships (and mediation is by definition a cause-
effect process), one must engage in experimental manipulation with
random assignment, collect data over time or, ideally, both.

Furthermore, one must meet an overwhelming number of

assumptions beyond those of linear modeling that go by such names
as “sequential ignorability,” “stable unit treatment value” and others,
many that are quite technical in nature or hard or impossible to test.

Others argue that one cannot conduct a mediation analysis with

merely correlational data, that moderators must be independent of
presumed causes of effects, and the list of requirements goes on and
on…”
 “Hardliners”

“(see e.g., Emsley, Dunn, & White, 2010; Preacher, 2015, for a
discussion of many of these assumptions).” “
 Hayes and Rockwood (2017)
“We feel that if these are taken as literal requirements rather than as
just ideals or recommendations, most research would not be done
because most researchers cannot meet these requirements (due to
resource constraints, ethics, and a myriad list of other reasons).”

“We would rather see more imperfect work conducted and published
than see research slow to a trickle because investigators don’t feel that
their work will satisfy all critics and pass every test for valid causal
inference.”

“You can do most anything you want with your data. Most any
statistical tool can provide some insight into the story you ultimately
end up telling with your data.”

Hayes and Rockwood (2017), BRAT, 98:39-57

 The Future for causal inference
• Training courses

• Masters level training

• European Causal Inference Meeting

– Bermen, April 2019

• Journal of Causal Inference

• Lots of forthcoming books

• Machine Learning and artificial intelligence

51
 Research Programme:
Efficacy and Mechanisms Evaluation
Joint work and slides prepared with Graham Dunn, Ian White, Andrew Pickles and Sabine Landau.

Funded by Medical Research Council Methodology Research Programmes:

• Design and methods of explanatory (causal) analysis for randomised trials of complex interventions in
mental health (2006-2009)
– Graham Dunn (PI), Richard Emsley, et al

• Estimation of causal effects of complex interventions in longitudinal studies with intermediate variables
(2009-2012)
– Richard Emsley (PI), Graham Dunn.

• MRC Early Career Centenary Award (2012-13)

• Designs and analysis for the evaluation and validation of social and psychological markers in randomised
trials of complex interventions in mental health (2010-12)
– Graham Dunn (PI), Richard Emsley, et al.

• Developing methods for understanding mechanism in complex interventions (2013-16)

– Sabine Landau (PI), Richard Emsley, et al.

• MRC NorthWest Hub for Trials Methodology Research (2013-2018)

– Paula Williamson (PI), Richard Emsley, et al.
52
 Methodology report
• Dunn G, Emsley RA, Liu H, Landau S, Green J, White I and
Pickles A. (2015). Evaluation and validation of social and
psychological markers in randomised trials of complex
interventions in mental health. Health Technology Assessment
19 (93).

• Non-technical introduction and summary of our work on

analysing complex interventions:
– Introduction to causal inference
– Mediation analysis
– Process evaluation
– Longitudinal extensions
– Stratified medicine
– Guidance and tips for trialists

53
Thank you for your attention

Email: richard.emsley@kcl.ac.uk

54